Junhan Yang

Kuan Lu, Shuhang Lin, Sai Wu et al.

Large language models (LLMs) are increasingly applied in long-context scenarios such as multi-turn conversations. However, long contexts pose significant challenges for inference efficiency, including high memory overhead from Key-Value (KV) cache and increased latency due to excessive memory accesses. Recent methods for dynamic KV selection struggle with trade-offs: block-level indexing degrades accuracy by retrieving irrelevant KV entries, while token-level indexing incurs high latency from inefficient retrieval mechanisms. In this paper, we propose CTKVR, a novel centroid-then-token KV retrieval scheme that addresses these limitations. CTKVR leverages a key observation: query vectors adjacent in position exhibit high similarity after Rotary Position Embedding (RoPE) and share most of their top-k KV cache entries. Based on this insight, CTKVR employs a two-stage retrieval strategy: lightweight centroids are precomputed during prefilling for centroid-grained indexing, followed by token-level refinement for precise KV retrieval. This approach balances retrieval efficiency and accuracy. To further enhance performance, we implement an optimized system for indexing construction and search using CPU-GPU co-execution. Experimentally, CTKVR achieves superior performance across multiple benchmarks with less than 1% accuracy degradation. Meanwhile, CTKVR delivers 3 times and 4 times throughput speedups on Llama-3-8B and Yi-9B at 96K context length across diverse GPU hardware.

Junhan Yang, Zheng Liu, Shitao Xiao et al.

The lack of labeled data is a major obstacle to learning high-quality sentence embeddings. Recently, self-supervised contrastive learning (SCL) is regarded as a promising way to address this problem. However, the existing works mainly rely on hand-crafted data annotation heuristics to generate positive training samples, which not only call for domain expertise and laborious tuning, but are also prone to the following unfavorable cases: 1) trivial positives, 2) coarse-grained positives, and 3) false positives. As a result, the self-supervision's quality can be severely limited in reality. In this work, we propose a novel framework InfoCSE to address the above problems. Instead of relying on annotation heuristics defined by humans, it leverages the sentence representation model itself and realizes the following iterative self-supervision process: on one hand, the improvement of sentence representation may contribute to the quality of data annotation; on the other hand, more effective data annotation helps to generate high-quality positive samples, which will further improve the current sentence representation model. In other words, the representation learning and data annotation become mutually reinforced, where a strong self-supervision effect can be derived. Extensive experiments are performed based on three benchmark datasets, where notable improvements can be achieved against the existing SCL-based methods.

Junhan Yang, Zheng Liu, Shitao Xiao et al.

The representation learning on textual graph is to generate low-dimensional embeddings for the nodes based on the individual textual features and the neighbourhood information. Recent breakthroughs on pretrained language models and graph neural networks push forward the development of corresponding techniques. The existing works mainly rely on the cascaded model architecture: the textual features of nodes are independently encoded by language models at first; the textual embeddings are aggregated by graph neural networks afterwards. However, the above architecture is limited due to the independent modeling of textual features. In this work, we propose GraphFormers, where layerwise GNN components are nested alongside the transformer blocks of language models. With the proposed architecture, the text encoding and the graph aggregation are fused into an iterative workflow, {making} each node's semantic accurately comprehended from the global perspective. In addition, a {progressive} learning strategy is introduced, where the model is successively trained on manipulated data and original data to reinforce its capability of integrating information on graph. Extensive evaluations are conducted on three large-scale benchmark datasets, where GraphFormers outperform the SOTA baselines with comparable running efficiency.

Junhan Yang, Zheng Liu, Bowen Jin et al.

Transformer encoding networks have been proved to be a powerful tool of understanding natural languages. They are playing a critical role in native ads service, which facilitates the recommendation of appropriate ads based on user's web browsing history. For the sake of efficient recommendation, conventional methods would generate user and advertisement embeddings independently with a siamese transformer encoder, such that approximate nearest neighbour search (ANN) can be leveraged. Given that the underlying semantic about user and ad can be complicated, such independently generated embeddings are prone to information loss, which leads to inferior recommendation quality. Although another encoding strategy, the cross encoder, can be much more accurate, it will lead to huge running cost and become infeasible for realtime services, like native ads recommendation. In this work, we propose hybrid encoder, which makes efficient and precise native ads recommendation through two consecutive steps: retrieval and ranking. In the retrieval step, user and ad are encoded with a siamese component, which enables relevant candidates to be retrieved via ANN search. In the ranking step, it further represents each ad with disentangled embeddings and each user with ad-related embeddings, which contributes to the fine-grained selection of high-quality ads from the candidate set. Both steps are light-weighted, thanks to the pre-computed and cached intermedia results. To optimize the hybrid encoder's performance in this two-stage workflow, a progressive training pipeline is developed, which builds up the model's capability in the retrieval and ranking task step-by-step. The hybrid encoder's effectiveness is experimentally verified: with very little additional cost, it outperforms the siamese encoder significantly and achieves comparable recommendation quality as the cross encoder.